*flush*

Seven local model/host combinations on an M5 Max with 128 GB, three passes each on the same agentic coding task, averaged: prefill, decode, memory, MTP acceptance, cache hit rate and code quality.

This guy is so well spoken. I'm a big fan of the positive changes AI has made in my life, and optimistic about what good things the future could hold. But we need to move slowly enough that we're designing systems and incentives that protect _people_, not IPOs.

Transparency note: this post was written by GLM 5.3 Flash working in the pi coding agent. There may be inaccuracies — verify against your own setup before relying on it. You can replace Jetpack's newsletter signup UI — the Subscribe block or the popup — with your own form, and keep driving the exact same WordPress.com subscription backend. This post explains how the integration works.

I'm lucky enough to have an M5 Max MacBook with 128gb of Ram, and I've been spending a lot of sweat this summer trying to understand the ins and outs of running local AI. And the ins and outs are a lot more detailed than I originally thought. There are lots of new little terms to learn, that are connected to big, complex concepts with major implications on how effectively a local model can run. I hope to put up more blog posts later with some details on the different setups I've been exploring.

Tool libraries look fantastic. I didn't know about them before watching this video but I looked them up and I have a local one that started this year. What a cool way to reduce consumer trash.

I spend so much time typing to AI bots these days that I get a sort of social vertigo when I'm on a support chat with a real human. I already practice being polite with bots because I like to, but when a human is on the other end I have to remember not to be as curt or short as I am with AI.

Kudos to Turo today. I had to cancel a non-refundable trip a few weeks from now, and they allowed me to ask the car’s owner if they were willing to do a full refund. The owner was very understanding, and they give me a full refund right away. Really boosts my willingness to use Turo again in the future.

A month ago I published a post about burning €11 in 18 minutes on GreenPT. The session logged 10 million input tokens with zero cache hits. Every request billed the full conversation history. GreenPT support responded to that post and refunded the €11. They said they do caching internally, but that it's opaque from the API — you wouldn't be able to see cache stats in the API response. That's no longer the case. Their pricing page now lists cached input rates for several models, and thecached_tokensfield shows up in the API response.

The detainee had an extensive medical history, including epilepsy, according to a lawsuit. He had complained about his face and hand going numb just before his death, his mother said.

Large changes are hard to review. Smaller pull requests make them easier to ship. Stacked PRs are now on GitHub. https://docs.github.com/en/pull-requests/how-tos/stacked-pull-requests?utm_source=bluesky-motion-graphic&utm_medium=social&utm_campaign=stacked-prs-gtm-public-preview-2026

Kalil's articles are important takes on history that the current US Gov't is trying to tear down in an attempt to preserve and consolidate their power. They are using the same tactics that other high-control cult-like groups use. Read people like Kalil to keep perspective:

The Endangered Species Act is wildly effective and widely popular. The law has prevented extinctions of numerous iconic species, recovered imperiled plants and animals, and protected the ecosystems we all rely on.

I've been trying to figure out how to make my AI usage less damaging to the environment. Most of my experiments so far have been about running local models on my MacBook, which uses way less power than an RTX card, or using small models like DeepSeek V4 Flash on OpenCode Go. Then I found GreenPT. They host smaller models in data centers that run on renewable energy and use less water for cooling. I made an account, loaded €25, and started a session with GLM-5.2 — a much stronger model than the ones I'd been using.

I'm building a side project — an offline-first PWA called Starchive where teams track appreciation with stars and hearts. Stars are admin-given points (gold, cashable). Hearts are anyone-to-anyone permanent thank-yous (pink, never cashed out). I'm doing this with opencode and opencode-go, using DeepSeek V4 Flash as my main model. It's fast and cheap — ideal for the 700+ message sessions where I'm iterating on code, and it helps me stay under my monthly usage limits. But here's the thing. DeepSeek V4 Flash is fast and cheap, and the UI it produces reflects that.

My previous post about Zanagrams.com made it sound like I’d built it. Sorry! I didn’t build it, just enjoyed it.

Beautiful words on freedom from Laura. Happy Independence Day to everyone in the world. May we all be free.

The kids and I had fun working on Zanagrams together today. It's like a Wordle, but different:

Whoops, looks like Crusoe’s reporting leaves something to be desired. I went far over my target billing limit before seeing any usage on the dashboard, and before being notified.

I've been playing with pointing my coding harness to Crusoe Cloud's models, since Crusoe harnesses otherwise overlooked energy sources. I'm surprised 😮 by how fast their version of Kimi K2.6 is ⚡

We need to be transitioning off of our dependence on big oil, not giving them new exceptions to extract profit from endangered nature. Please weigh in, either by calling your reps or using this site:

We absolutely need to retain the right to sue big oil companies. Evidently US senators are trying to give them immunity. If you're a US citizen, please contact your reps to stop it:

California is experimenting with covering its human-made waterways with solar panels, like other countries have done. It could be huge both for renewable energy and for water conservation. I’m hoping there aren’t nasty side effects on wildlife.

Claude's over-use of the expression "is real" is real.

The other day while waiting for Claude to process prompts, I watched this video: (I know, the irony isn't lost on me) It highlights the value of reducing multitasking and freeing up brain space. At this point I'm not going to radically change my routine the way Lucas did. But here's something I am trying; giving command-tab a break. Between prompts in claude code, I noticed my self tabbing away instantly to read messages in Slack, to check Reddit for news, to send a different prompt in another window, or (ironically) to write this blog post.

In the AI age, we need GitHub to allow LLMs to upload screenshots to pull requests without weird workarounds. This morning Claude was working in a repo that didn't have screenshot upload skills added. When it had trouble attaching screenshots to a PR comment, it went ahead and uploaded those screenshots to imgur.com without consulting me first. That's a big-old yikes 😂😱

argh, I'm sad that yet another tool I want to pay for is going to be putting money in Musk's pocket. I was leaning on Cursor for personal work, but it's time to find something else.

According to Reddit, horrible people have made a nasty supply-chain attack called Hades: Looks like it's marked at this time by: the presence of suspicious *-setup.pth files in your python site-packages A suspicious .bun_ran marker in your tmp dirs Weird instructions in your agent harnesses or config files If you find any of these things, you'll want to disable your network connection immediately, clean them up fully, and then rotate any and all credentials your computer had on-disk. Looking for trouble You can ask your agent to look for these things for you, but it's possible that malicious instructions could alter the agent's behavior to try to conceal the bad stuff, so here's a script you can read yourself (hah, as if Bash were readable) to check for the symptomatic files in site-packages and tmp, and to give you a list of agent config files for you to check manually for "weird instructions".

When there's this much money and power at stake, I suspect there aren't really any big companies that are the "good guys", but when Anthropic is spending money and attention on things like this, it does a lot to sway my opinion:

I want to trust AI with code enough that I'm no longer doing line-by-line reviews. But when I look at the code, even from frontier models, I'm still finding small things that feel smelly. In a simple PR Sonnet 4.6 made, I just found it changing the default mock data constructor for all existing test cases away from the common case to a less-common scenario. All the existing tests passed because they asserted on things that weren't changed. The update was technically sound, but it's the kind of change that would feel weird to a human if they came across it later.

Diving into test assertions is a good way to focus on the verifiability of a PR when doing AI-assisted reviews. Here's one of the prompts I often use: please give me a human-readable outline of thetests we added specifically calling out what theyassert and how as an html doc

I'm slowly watching, feeling out the shape of this wave that we're all riding together. I'm sensing an order-of-magnitude zoom-out when it comes to the relationship between our brains and the code we manage. My colleagues, while working to understand our changing roles, have drawn comparisons between AI-produced code and compilers. Most programmers stopped paying attention to Assembly code when compilers became reliable, and reasoned only on the level of the language code. Could it be that AI models are the next step in that evolution? Will we stop reading the code soon?

Yesterday, like many of you, I read about Jack Dorsey announcing the end of employment for 4,000 people at Blocks after a strong year of profit increases. It's another splash in the pool of anxiety engineers and creatives alike are swimming in as we watch the astoundingly fast evolution of AI agents. Rumination How do we process this? Thousands of our career-siblings loosing the employment that our society is structured to make non-optional with a well-meaning "good luck" pat on the back from the people who just hired them. If you feel offended at this statement, and move to defend it as "the way things go in business", I understand.

AI autonomously manages a theme park in the classic game RollerCoaster Tycoon, placing rides, fixing infrastructure, and generating CFO reports, all via command line.

Code written with AI agents gets substantially more robust when the agent has a feedback loop, like a good test suite. And agents like Claude Code are very good at writing big.. long... super long... test suites. I find myself not only generating more application code than I've ever written in my life, but many more lines of testing code.

I've been experimenting with using Claude Code to create Linear issues instead of filling them out by hand, and it's been a bigger workflow upgrade than I’d expected it to be. Too often I’ve made or come across Linear issues that are just a line or two, jotted down from someone’s memory and now turned inscrutable. Having Claude flesh out the details is a relief, especially when it then includes the Linear issue ID in its branch, so that the PR it produces gets automatically linked.

My attention wanders. A lot. I'll kick off a Claude Code session, switch to check on something else, and 20 minutes later realize I've got a forlorn Claude tab that’s been waiting for approval for …19 minutes. A notification beep doesn't help much—it starts a quick game of "which tab was that?”. Not a huge time sink, since it takes only a moment to find the tab.

StopNCII.org is operated by the Revenge Porn Helpline which is part of SWGfL, a charity that believes that all should benefit from technology, free from harm.